| Recognizing and modeling social behaviors of animals has many applications, including: (1) improved understanding of their behavior; (2) enhanced protection of species; (3) hosting of specimens in enriched environment at zoos; and (4) efficient extraction and analysis of data in important basic and applied biological research, where animal models are used. Currently, understanding social behaviors in animals is achieved either by direct human observations or by videotaping and then coding the behaviors. Both of these approaches have major limitations including being heavily time consuming and requiring highly trained behavioral science experts. Having an automated system to recognize and model social behaviors would facilitate the scientific study of complex behaviors with less impact due to these constraints. However, research in this area is very limited.;In this dissertation, we describe a framework that adopts current practices from computer vision and machine learning in creating the preliminary steps towards solving the problem of automatically recognizing behaviors of primates in a social group (in this case, a pen hosting a group of 3 or more primates). Several challenges need to be overcome in order to achieve primate activity recognition from videos, some of which are: the massive size of continuous video recordings from multiple cameras over days and weeks, illumination variations throughout the day, background changes due to moving objects in the pen and humans passing by (e.g. for feeding or observing), highly variable shapes and poses of primates, and the low visibility of color-coded primate collars causing difficulty in identifying the primates.;This study is unique, to our knowledge, because it tackles automatic primate behavior and interaction recognition in social groups hosted in a pen for the first time. Results indicate that the activities extracted based on the detection and tracking algorithms developed are sufficiently accurate to infer primate behaviors and social interactions. |
